Recommender systems have become increasingly important in our daily lives since they play an important role in mitigating the information overload problem, especially in many user-oriented online services such as E-commerce (Amazon, Taobao), and Social Media Sites (Facebook, Instagram). The goal of recommender systems is to suggest a personalized list of items that users are likely to click or purchase. Among existing techniques in modern recommender systems, collaborative filtering (CF) is one of the most popular techniques to model users' preferences towards items by utilizing the history of user-item interactions.

In addition to the user-item interactions, social relations among users provide another stream of potential information to understand users' preferences. More specifically, users are able to acquire and disseminate information through people around them, such as classmates, friends, relatives, or colleagues. The objective of this thesis is to explore approaches to utilize users' social relations to boost the performance of recommender systems.

In this thesis, we aim to build social recommender systems based on deep neural networks. More specifically, we propose three novel deep neural networks for social recommendations, which can model social relations from various perspectives for recommender systems. The majority of existing social recommender systems only involve direct social neighbors, while information from users who are a few hops away could also be helpful. Thus, we first propose a Deep Social Collaborative Filtering framework (DSCF), which adopts the Bi-directional Long Short-Term Memory Network (Bi-LSTM) with attention mechanism to extract useful information from distant neighbors. Moreover, data in social recommendation can be naturally represented as graph data with two explicit graphs, a user-item graph denoting interactions between users and items, and a social graph denoting the relationships between users. These two types of user relationships can help infer better user preferences from different perspectives. Therefore, we propose GraphRec, which aims to build social recommender systems based on graph neural networks. GraphRec is proposed to coherently model graph data to learn better user and item representations for social recommendation. At last, using a unified user representation in social recommendation may restrain user representation learning, and results in an inflexible/limited transferring of knowledge from the social relations for recommendation. Meanwhile, utilizing negative sampling technique to optimize the user/item representations is quite ineffective. Therefore, we propose a novel deep adversarial social recommendation framework (DASO), which aims to learn separated user representations in the two domains with adversarial learning. The experimental results show that our proposed social recommender systems can achieve promising performance on real-world datasets.